xref: /freebsd/sys/vm/vm_pageout.c (revision dadef94c7a762d05890e2891bc4a7d1dfe0cf758)
1 /*-
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1994 John S. Dyson
5  * All rights reserved.
6  * Copyright (c) 1994 David Greenman
7  * All rights reserved.
8  * Copyright (c) 2005 Yahoo! Technologies Norway AS
9  * All rights reserved.
10  *
11  * This code is derived from software contributed to Berkeley by
12  * The Mach Operating System project at Carnegie-Mellon University.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 3. All advertising materials mentioning features or use of this software
23  *    must display the following acknowledgement:
24  *	This product includes software developed by the University of
25  *	California, Berkeley and its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
43  *
44  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49  *
50  * Permission to use, copy, modify and distribute this software and
51  * its documentation is hereby granted, provided that both the copyright
52  * notice and this permission notice appear in all copies of the
53  * software, derivative works or modified versions, and any portions
54  * thereof, and that both notices appear in supporting documentation.
55  *
56  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59  *
60  * Carnegie Mellon requests users of this software to return to
61  *
62  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
63  *  School of Computer Science
64  *  Carnegie Mellon University
65  *  Pittsburgh PA 15213-3890
66  *
67  * any improvements or extensions that they make and grant Carnegie the
68  * rights to redistribute these changes.
69  */
70 
71 /*
72  *	The proverbial page-out daemon.
73  */
74 
75 #include <sys/cdefs.h>
76 __FBSDID("$FreeBSD$");
77 
78 #include "opt_vm.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/eventhandler.h>
83 #include <sys/lock.h>
84 #include <sys/mutex.h>
85 #include <sys/proc.h>
86 #include <sys/kthread.h>
87 #include <sys/ktr.h>
88 #include <sys/mount.h>
89 #include <sys/resourcevar.h>
90 #include <sys/sched.h>
91 #include <sys/signalvar.h>
92 #include <sys/vnode.h>
93 #include <sys/vmmeter.h>
94 #include <sys/sx.h>
95 #include <sys/sysctl.h>
96 
97 #include <vm/vm.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_pager.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_extern.h>
106 #include <vm/uma.h>
107 
108 /*
109  * System initialization
110  */
111 
112 /* the kernel process "vm_pageout"*/
113 static void vm_pageout(void);
114 static int vm_pageout_clean(vm_page_t);
115 static void vm_pageout_scan(int pass);
116 
117 struct proc *pageproc;
118 
119 static struct kproc_desc page_kp = {
120 	"pagedaemon",
121 	vm_pageout,
122 	&pageproc
123 };
124 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
125     &page_kp);
126 
127 #if !defined(NO_SWAPPING)
128 /* the kernel process "vm_daemon"*/
129 static void vm_daemon(void);
130 static struct	proc *vmproc;
131 
132 static struct kproc_desc vm_kp = {
133 	"vmdaemon",
134 	vm_daemon,
135 	&vmproc
136 };
137 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
138 #endif
139 
140 
141 int vm_pages_needed;		/* Event on which pageout daemon sleeps */
142 int vm_pageout_deficit;		/* Estimated number of pages deficit */
143 int vm_pageout_pages_needed;	/* flag saying that the pageout daemon needs pages */
144 
145 #if !defined(NO_SWAPPING)
146 static int vm_pageout_req_swapout;	/* XXX */
147 static int vm_daemon_needed;
148 static struct mtx vm_daemon_mtx;
149 /* Allow for use by vm_pageout before vm_daemon is initialized. */
150 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
151 #endif
152 static int vm_max_launder = 32;
153 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
154 static int vm_pageout_full_stats_interval = 0;
155 static int vm_pageout_algorithm=0;
156 static int defer_swap_pageouts=0;
157 static int disable_swap_pageouts=0;
158 
159 #if defined(NO_SWAPPING)
160 static int vm_swap_enabled=0;
161 static int vm_swap_idle_enabled=0;
162 #else
163 static int vm_swap_enabled=1;
164 static int vm_swap_idle_enabled=0;
165 #endif
166 
167 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
168 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
169 
170 SYSCTL_INT(_vm, OID_AUTO, max_launder,
171 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
172 
173 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
174 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
175 
176 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
177 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
178 
179 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
180 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
181 
182 #if defined(NO_SWAPPING)
183 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
184 	CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
185 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
186 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
187 #else
188 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
189 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
190 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
191 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
192 #endif
193 
194 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
195 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
196 
197 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
198 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
199 
200 static int pageout_lock_miss;
201 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
202 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
203 
204 #define VM_PAGEOUT_PAGE_COUNT 16
205 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
206 
207 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
208 SYSCTL_INT(_vm, OID_AUTO, max_wired,
209 	CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
210 
211 #if !defined(NO_SWAPPING)
212 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
213 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
214 static void vm_req_vmdaemon(int req);
215 #endif
216 static void vm_pageout_page_stats(void);
217 
218 static void
219 vm_pageout_init_marker(vm_page_t marker, u_short queue)
220 {
221 
222 	bzero(marker, sizeof(*marker));
223 	marker->flags = PG_FICTITIOUS | PG_MARKER;
224 	marker->oflags = VPO_BUSY;
225 	marker->queue = queue;
226 	marker->wire_count = 1;
227 }
228 
229 /*
230  * vm_pageout_fallback_object_lock:
231  *
232  * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
233  * known to have failed and page queue must be either PQ_ACTIVE or
234  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
235  * while locking the vm object.  Use marker page to detect page queue
236  * changes and maintain notion of next page on page queue.  Return
237  * TRUE if no changes were detected, FALSE otherwise.  vm object is
238  * locked on return.
239  *
240  * This function depends on both the lock portion of struct vm_object
241  * and normal struct vm_page being type stable.
242  */
243 boolean_t
244 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
245 {
246 	struct vm_page marker;
247 	boolean_t unchanged;
248 	u_short queue;
249 	vm_object_t object;
250 
251 	queue = m->queue;
252 	vm_pageout_init_marker(&marker, queue);
253 	object = m->object;
254 
255 	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
256 			   m, &marker, pageq);
257 	vm_page_unlock_queues();
258 	vm_page_unlock(m);
259 	VM_OBJECT_LOCK(object);
260 	vm_page_lock(m);
261 	vm_page_lock_queues();
262 
263 	/* Page queue might have changed. */
264 	*next = TAILQ_NEXT(&marker, pageq);
265 	unchanged = (m->queue == queue &&
266 		     m->object == object &&
267 		     &marker == TAILQ_NEXT(m, pageq));
268 	TAILQ_REMOVE(&vm_page_queues[queue].pl,
269 		     &marker, pageq);
270 	return (unchanged);
271 }
272 
273 /*
274  * Lock the page while holding the page queue lock.  Use marker page
275  * to detect page queue changes and maintain notion of next page on
276  * page queue.  Return TRUE if no changes were detected, FALSE
277  * otherwise.  The page is locked on return. The page queue lock might
278  * be dropped and reacquired.
279  *
280  * This function depends on normal struct vm_page being type stable.
281  */
282 boolean_t
283 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
284 {
285 	struct vm_page marker;
286 	boolean_t unchanged;
287 	u_short queue;
288 
289 	vm_page_lock_assert(m, MA_NOTOWNED);
290 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
291 
292 	if (vm_page_trylock(m))
293 		return (TRUE);
294 
295 	queue = m->queue;
296 	vm_pageout_init_marker(&marker, queue);
297 
298 	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, m, &marker, pageq);
299 	vm_page_unlock_queues();
300 	vm_page_lock(m);
301 	vm_page_lock_queues();
302 
303 	/* Page queue might have changed. */
304 	*next = TAILQ_NEXT(&marker, pageq);
305 	unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, pageq));
306 	TAILQ_REMOVE(&vm_page_queues[queue].pl, &marker, pageq);
307 	return (unchanged);
308 }
309 
310 /*
311  * vm_pageout_clean:
312  *
313  * Clean the page and remove it from the laundry.
314  *
315  * We set the busy bit to cause potential page faults on this page to
316  * block.  Note the careful timing, however, the busy bit isn't set till
317  * late and we cannot do anything that will mess with the page.
318  */
319 static int
320 vm_pageout_clean(vm_page_t m)
321 {
322 	vm_object_t object;
323 	vm_page_t mc[2*vm_pageout_page_count], pb, ps;
324 	int pageout_count;
325 	int ib, is, page_base;
326 	vm_pindex_t pindex = m->pindex;
327 
328 	vm_page_lock_assert(m, MA_OWNED);
329 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
330 
331 	/*
332 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
333 	 * with the new swapper, but we could have serious problems paging
334 	 * out other object types if there is insufficient memory.
335 	 *
336 	 * Unfortunately, checking free memory here is far too late, so the
337 	 * check has been moved up a procedural level.
338 	 */
339 
340 	/*
341 	 * Can't clean the page if it's busy or held.
342 	 */
343 	KASSERT(m->busy == 0 && (m->oflags & VPO_BUSY) == 0,
344 	    ("vm_pageout_clean: page %p is busy", m));
345 	KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
346 
347 	mc[vm_pageout_page_count] = pb = ps = m;
348 	pageout_count = 1;
349 	page_base = vm_pageout_page_count;
350 	ib = 1;
351 	is = 1;
352 
353 	/*
354 	 * Scan object for clusterable pages.
355 	 *
356 	 * We can cluster ONLY if: ->> the page is NOT
357 	 * clean, wired, busy, held, or mapped into a
358 	 * buffer, and one of the following:
359 	 * 1) The page is inactive, or a seldom used
360 	 *    active page.
361 	 * -or-
362 	 * 2) we force the issue.
363 	 *
364 	 * During heavy mmap/modification loads the pageout
365 	 * daemon can really fragment the underlying file
366 	 * due to flushing pages out of order and not trying
367 	 * align the clusters (which leave sporatic out-of-order
368 	 * holes).  To solve this problem we do the reverse scan
369 	 * first and attempt to align our cluster, then do a
370 	 * forward scan if room remains.
371 	 */
372 	object = m->object;
373 more:
374 	while (ib && pageout_count < vm_pageout_page_count) {
375 		vm_page_t p;
376 
377 		if (ib > pindex) {
378 			ib = 0;
379 			break;
380 		}
381 
382 		if ((p = vm_page_prev(pb)) == NULL ||
383 		    (p->oflags & VPO_BUSY) != 0 || p->busy != 0) {
384 			ib = 0;
385 			break;
386 		}
387 		vm_page_lock(p);
388 		vm_page_test_dirty(p);
389 		if (p->dirty == 0 ||
390 		    p->queue != PQ_INACTIVE ||
391 		    p->hold_count != 0) {	/* may be undergoing I/O */
392 			vm_page_unlock(p);
393 			ib = 0;
394 			break;
395 		}
396 		vm_page_unlock(p);
397 		mc[--page_base] = pb = p;
398 		++pageout_count;
399 		++ib;
400 		/*
401 		 * alignment boundry, stop here and switch directions.  Do
402 		 * not clear ib.
403 		 */
404 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
405 			break;
406 	}
407 
408 	while (pageout_count < vm_pageout_page_count &&
409 	    pindex + is < object->size) {
410 		vm_page_t p;
411 
412 		if ((p = vm_page_next(ps)) == NULL ||
413 		    (p->oflags & VPO_BUSY) != 0 || p->busy != 0)
414 			break;
415 		vm_page_lock(p);
416 		vm_page_test_dirty(p);
417 		if (p->dirty == 0 ||
418 		    p->queue != PQ_INACTIVE ||
419 		    p->hold_count != 0) {	/* may be undergoing I/O */
420 			vm_page_unlock(p);
421 			break;
422 		}
423 		vm_page_unlock(p);
424 		mc[page_base + pageout_count] = ps = p;
425 		++pageout_count;
426 		++is;
427 	}
428 
429 	/*
430 	 * If we exhausted our forward scan, continue with the reverse scan
431 	 * when possible, even past a page boundry.  This catches boundry
432 	 * conditions.
433 	 */
434 	if (ib && pageout_count < vm_pageout_page_count)
435 		goto more;
436 
437 	vm_page_unlock(m);
438 	/*
439 	 * we allow reads during pageouts...
440 	 */
441 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
442 }
443 
444 /*
445  * vm_pageout_flush() - launder the given pages
446  *
447  *	The given pages are laundered.  Note that we setup for the start of
448  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
449  *	reference count all in here rather then in the parent.  If we want
450  *	the parent to do more sophisticated things we may have to change
451  *	the ordering.
452  */
453 int
454 vm_pageout_flush(vm_page_t *mc, int count, int flags)
455 {
456 	vm_object_t object = mc[0]->object;
457 	int pageout_status[count];
458 	int numpagedout = 0;
459 	int i;
460 
461 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
462 	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
463 
464 	/*
465 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
466 	 * mark the pages read-only.
467 	 *
468 	 * We do not have to fixup the clean/dirty bits here... we can
469 	 * allow the pager to do it after the I/O completes.
470 	 *
471 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
472 	 * edge case with file fragments.
473 	 */
474 	for (i = 0; i < count; i++) {
475 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
476 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
477 			mc[i], i, count));
478 		vm_page_io_start(mc[i]);
479 		pmap_remove_write(mc[i]);
480 	}
481 	vm_object_pip_add(object, count);
482 
483 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
484 
485 	for (i = 0; i < count; i++) {
486 		vm_page_t mt = mc[i];
487 
488 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
489 		    (mt->flags & PG_WRITEABLE) == 0,
490 		    ("vm_pageout_flush: page %p is not write protected", mt));
491 		switch (pageout_status[i]) {
492 		case VM_PAGER_OK:
493 		case VM_PAGER_PEND:
494 			numpagedout++;
495 			break;
496 		case VM_PAGER_BAD:
497 			/*
498 			 * Page outside of range of object. Right now we
499 			 * essentially lose the changes by pretending it
500 			 * worked.
501 			 */
502 			vm_page_undirty(mt);
503 			break;
504 		case VM_PAGER_ERROR:
505 		case VM_PAGER_FAIL:
506 			/*
507 			 * If page couldn't be paged out, then reactivate the
508 			 * page so it doesn't clog the inactive list.  (We
509 			 * will try paging out it again later).
510 			 */
511 			vm_page_lock(mt);
512 			vm_page_activate(mt);
513 			vm_page_unlock(mt);
514 			break;
515 		case VM_PAGER_AGAIN:
516 			break;
517 		}
518 
519 		/*
520 		 * If the operation is still going, leave the page busy to
521 		 * block all other accesses. Also, leave the paging in
522 		 * progress indicator set so that we don't attempt an object
523 		 * collapse.
524 		 */
525 		if (pageout_status[i] != VM_PAGER_PEND) {
526 			vm_object_pip_wakeup(object);
527 			vm_page_io_finish(mt);
528 			if (vm_page_count_severe()) {
529 				vm_page_lock(mt);
530 				vm_page_try_to_cache(mt);
531 				vm_page_unlock(mt);
532 			}
533 		}
534 	}
535 	return (numpagedout);
536 }
537 
538 #if !defined(NO_SWAPPING)
539 /*
540  *	vm_pageout_object_deactivate_pages
541  *
542  *	Deactivate enough pages to satisfy the inactive target
543  *	requirements.
544  *
545  *	The object and map must be locked.
546  */
547 static void
548 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
549     long desired)
550 {
551 	vm_object_t backing_object, object;
552 	vm_page_t p;
553 	int actcount, remove_mode;
554 
555 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
556 	if (first_object->type == OBJT_DEVICE ||
557 	    first_object->type == OBJT_SG)
558 		return;
559 	for (object = first_object;; object = backing_object) {
560 		if (pmap_resident_count(pmap) <= desired)
561 			goto unlock_return;
562 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
563 		if (object->type == OBJT_PHYS || object->paging_in_progress)
564 			goto unlock_return;
565 
566 		remove_mode = 0;
567 		if (object->shadow_count > 1)
568 			remove_mode = 1;
569 		/*
570 		 * Scan the object's entire memory queue.
571 		 */
572 		TAILQ_FOREACH(p, &object->memq, listq) {
573 			if (pmap_resident_count(pmap) <= desired)
574 				goto unlock_return;
575 			if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0)
576 				continue;
577 			PCPU_INC(cnt.v_pdpages);
578 			vm_page_lock(p);
579 			if (p->wire_count != 0 || p->hold_count != 0 ||
580 			    !pmap_page_exists_quick(pmap, p)) {
581 				vm_page_unlock(p);
582 				continue;
583 			}
584 			actcount = pmap_ts_referenced(p);
585 			if ((p->flags & PG_REFERENCED) != 0) {
586 				if (actcount == 0)
587 					actcount = 1;
588 				vm_page_lock_queues();
589 				vm_page_flag_clear(p, PG_REFERENCED);
590 				vm_page_unlock_queues();
591 			}
592 			if (p->queue != PQ_ACTIVE && actcount != 0) {
593 				vm_page_activate(p);
594 				p->act_count += actcount;
595 			} else if (p->queue == PQ_ACTIVE) {
596 				if (actcount == 0) {
597 					p->act_count -= min(p->act_count,
598 					    ACT_DECLINE);
599 					if (!remove_mode &&
600 					    (vm_pageout_algorithm ||
601 					    p->act_count == 0)) {
602 						pmap_remove_all(p);
603 						vm_page_deactivate(p);
604 					} else {
605 						vm_page_lock_queues();
606 						vm_page_requeue(p);
607 						vm_page_unlock_queues();
608 					}
609 				} else {
610 					vm_page_activate(p);
611 					if (p->act_count < ACT_MAX -
612 					    ACT_ADVANCE)
613 						p->act_count += ACT_ADVANCE;
614 					vm_page_lock_queues();
615 					vm_page_requeue(p);
616 					vm_page_unlock_queues();
617 				}
618 			} else if (p->queue == PQ_INACTIVE)
619 				pmap_remove_all(p);
620 			vm_page_unlock(p);
621 		}
622 		if ((backing_object = object->backing_object) == NULL)
623 			goto unlock_return;
624 		VM_OBJECT_LOCK(backing_object);
625 		if (object != first_object)
626 			VM_OBJECT_UNLOCK(object);
627 	}
628 unlock_return:
629 	if (object != first_object)
630 		VM_OBJECT_UNLOCK(object);
631 }
632 
633 /*
634  * deactivate some number of pages in a map, try to do it fairly, but
635  * that is really hard to do.
636  */
637 static void
638 vm_pageout_map_deactivate_pages(map, desired)
639 	vm_map_t map;
640 	long desired;
641 {
642 	vm_map_entry_t tmpe;
643 	vm_object_t obj, bigobj;
644 	int nothingwired;
645 
646 	if (!vm_map_trylock(map))
647 		return;
648 
649 	bigobj = NULL;
650 	nothingwired = TRUE;
651 
652 	/*
653 	 * first, search out the biggest object, and try to free pages from
654 	 * that.
655 	 */
656 	tmpe = map->header.next;
657 	while (tmpe != &map->header) {
658 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
659 			obj = tmpe->object.vm_object;
660 			if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
661 				if (obj->shadow_count <= 1 &&
662 				    (bigobj == NULL ||
663 				     bigobj->resident_page_count < obj->resident_page_count)) {
664 					if (bigobj != NULL)
665 						VM_OBJECT_UNLOCK(bigobj);
666 					bigobj = obj;
667 				} else
668 					VM_OBJECT_UNLOCK(obj);
669 			}
670 		}
671 		if (tmpe->wired_count > 0)
672 			nothingwired = FALSE;
673 		tmpe = tmpe->next;
674 	}
675 
676 	if (bigobj != NULL) {
677 		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
678 		VM_OBJECT_UNLOCK(bigobj);
679 	}
680 	/*
681 	 * Next, hunt around for other pages to deactivate.  We actually
682 	 * do this search sort of wrong -- .text first is not the best idea.
683 	 */
684 	tmpe = map->header.next;
685 	while (tmpe != &map->header) {
686 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
687 			break;
688 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
689 			obj = tmpe->object.vm_object;
690 			if (obj != NULL) {
691 				VM_OBJECT_LOCK(obj);
692 				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
693 				VM_OBJECT_UNLOCK(obj);
694 			}
695 		}
696 		tmpe = tmpe->next;
697 	}
698 
699 	/*
700 	 * Remove all mappings if a process is swapped out, this will free page
701 	 * table pages.
702 	 */
703 	if (desired == 0 && nothingwired) {
704 		tmpe = map->header.next;
705 		while (tmpe != &map->header) {
706 			pmap_remove(vm_map_pmap(map), tmpe->start, tmpe->end);
707 			tmpe = tmpe->next;
708 		}
709 	}
710 	vm_map_unlock(map);
711 }
712 #endif		/* !defined(NO_SWAPPING) */
713 
714 /*
715  *	vm_pageout_scan does the dirty work for the pageout daemon.
716  */
717 static void
718 vm_pageout_scan(int pass)
719 {
720 	vm_page_t m, next;
721 	struct vm_page marker;
722 	int page_shortage, maxscan, pcount;
723 	int addl_page_shortage, addl_page_shortage_init;
724 	vm_object_t object;
725 	int actcount;
726 	int vnodes_skipped = 0;
727 	int maxlaunder;
728 
729 	/*
730 	 * Decrease registered cache sizes.
731 	 */
732 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
733 	/*
734 	 * We do this explicitly after the caches have been drained above.
735 	 */
736 	uma_reclaim();
737 
738 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
739 
740 	/*
741 	 * Calculate the number of pages we want to either free or move
742 	 * to the cache.
743 	 */
744 	page_shortage = vm_paging_target() + addl_page_shortage_init;
745 
746 	vm_pageout_init_marker(&marker, PQ_INACTIVE);
747 
748 	/*
749 	 * Start scanning the inactive queue for pages we can move to the
750 	 * cache or free.  The scan will stop when the target is reached or
751 	 * we have scanned the entire inactive queue.  Note that m->act_count
752 	 * is not used to form decisions for the inactive queue, only for the
753 	 * active queue.
754 	 *
755 	 * maxlaunder limits the number of dirty pages we flush per scan.
756 	 * For most systems a smaller value (16 or 32) is more robust under
757 	 * extreme memory and disk pressure because any unnecessary writes
758 	 * to disk can result in extreme performance degredation.  However,
759 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
760 	 * used) will die horribly with limited laundering.  If the pageout
761 	 * daemon cannot clean enough pages in the first pass, we let it go
762 	 * all out in succeeding passes.
763 	 */
764 	if ((maxlaunder = vm_max_launder) <= 1)
765 		maxlaunder = 1;
766 	if (pass)
767 		maxlaunder = 10000;
768 	vm_page_lock_queues();
769 rescan0:
770 	addl_page_shortage = addl_page_shortage_init;
771 	maxscan = cnt.v_inactive_count;
772 
773 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
774 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
775 	     m = next) {
776 
777 		cnt.v_pdpages++;
778 
779 		if (m->queue != PQ_INACTIVE)
780 			goto rescan0;
781 
782 		next = TAILQ_NEXT(m, pageq);
783 
784 		/*
785 		 * skip marker pages
786 		 */
787 		if (m->flags & PG_MARKER)
788 			continue;
789 
790 		/*
791 		 * Lock the page.
792 		 */
793 		if (!vm_pageout_page_lock(m, &next)) {
794 			vm_page_unlock(m);
795 			addl_page_shortage++;
796 			continue;
797 		}
798 
799 		/*
800 		 * A held page may be undergoing I/O, so skip it.
801 		 */
802 		if (m->hold_count) {
803 			vm_page_unlock(m);
804 			vm_page_requeue(m);
805 			addl_page_shortage++;
806 			continue;
807 		}
808 
809 		/*
810 		 * Don't mess with busy pages, keep in the front of the
811 		 * queue, most likely are being paged out.
812 		 */
813 		object = m->object;
814 		if (!VM_OBJECT_TRYLOCK(object) &&
815 		    (!vm_pageout_fallback_object_lock(m, &next) ||
816 			m->hold_count != 0)) {
817 			VM_OBJECT_UNLOCK(object);
818 			vm_page_unlock(m);
819 			addl_page_shortage++;
820 			continue;
821 		}
822 		if (m->busy || (m->oflags & VPO_BUSY)) {
823 			vm_page_unlock(m);
824 			VM_OBJECT_UNLOCK(object);
825 			addl_page_shortage++;
826 			continue;
827 		}
828 
829 		/*
830 		 * If the object is not being used, we ignore previous
831 		 * references.
832 		 */
833 		if (object->ref_count == 0) {
834 			vm_page_flag_clear(m, PG_REFERENCED);
835 			KASSERT(!pmap_page_is_mapped(m),
836 			    ("vm_pageout_scan: page %p is mapped", m));
837 
838 		/*
839 		 * Otherwise, if the page has been referenced while in the
840 		 * inactive queue, we bump the "activation count" upwards,
841 		 * making it less likely that the page will be added back to
842 		 * the inactive queue prematurely again.  Here we check the
843 		 * page tables (or emulated bits, if any), given the upper
844 		 * level VM system not knowing anything about existing
845 		 * references.
846 		 */
847 		} else if (((m->flags & PG_REFERENCED) == 0) &&
848 			(actcount = pmap_ts_referenced(m))) {
849 			vm_page_activate(m);
850 			VM_OBJECT_UNLOCK(object);
851 			m->act_count += (actcount + ACT_ADVANCE);
852 			vm_page_unlock(m);
853 			continue;
854 		}
855 
856 		/*
857 		 * If the upper level VM system knows about any page
858 		 * references, we activate the page.  We also set the
859 		 * "activation count" higher than normal so that we will less
860 		 * likely place pages back onto the inactive queue again.
861 		 */
862 		if ((m->flags & PG_REFERENCED) != 0) {
863 			vm_page_flag_clear(m, PG_REFERENCED);
864 			actcount = pmap_ts_referenced(m);
865 			vm_page_activate(m);
866 			VM_OBJECT_UNLOCK(object);
867 			m->act_count += (actcount + ACT_ADVANCE + 1);
868 			vm_page_unlock(m);
869 			continue;
870 		}
871 
872 		/*
873 		 * If the upper level VM system does not believe that the page
874 		 * is fully dirty, but it is mapped for write access, then we
875 		 * consult the pmap to see if the page's dirty status should
876 		 * be updated.
877 		 */
878 		if (m->dirty != VM_PAGE_BITS_ALL &&
879 		    (m->flags & PG_WRITEABLE) != 0) {
880 			/*
881 			 * Avoid a race condition: Unless write access is
882 			 * removed from the page, another processor could
883 			 * modify it before all access is removed by the call
884 			 * to vm_page_cache() below.  If vm_page_cache() finds
885 			 * that the page has been modified when it removes all
886 			 * access, it panics because it cannot cache dirty
887 			 * pages.  In principle, we could eliminate just write
888 			 * access here rather than all access.  In the expected
889 			 * case, when there are no last instant modifications
890 			 * to the page, removing all access will be cheaper
891 			 * overall.
892 			 */
893 			if (pmap_is_modified(m))
894 				vm_page_dirty(m);
895 			else if (m->dirty == 0)
896 				pmap_remove_all(m);
897 		}
898 
899 		if (m->valid == 0) {
900 			/*
901 			 * Invalid pages can be easily freed
902 			 */
903 			vm_page_free(m);
904 			cnt.v_dfree++;
905 			--page_shortage;
906 		} else if (m->dirty == 0) {
907 			/*
908 			 * Clean pages can be placed onto the cache queue.
909 			 * This effectively frees them.
910 			 */
911 			vm_page_cache(m);
912 			--page_shortage;
913 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
914 			/*
915 			 * Dirty pages need to be paged out, but flushing
916 			 * a page is extremely expensive verses freeing
917 			 * a clean page.  Rather then artificially limiting
918 			 * the number of pages we can flush, we instead give
919 			 * dirty pages extra priority on the inactive queue
920 			 * by forcing them to be cycled through the queue
921 			 * twice before being flushed, after which the
922 			 * (now clean) page will cycle through once more
923 			 * before being freed.  This significantly extends
924 			 * the thrash point for a heavily loaded machine.
925 			 */
926 			vm_page_flag_set(m, PG_WINATCFLS);
927 			vm_page_requeue(m);
928 		} else if (maxlaunder > 0) {
929 			/*
930 			 * We always want to try to flush some dirty pages if
931 			 * we encounter them, to keep the system stable.
932 			 * Normally this number is small, but under extreme
933 			 * pressure where there are insufficient clean pages
934 			 * on the inactive queue, we may have to go all out.
935 			 */
936 			int swap_pageouts_ok, vfslocked = 0;
937 			struct vnode *vp = NULL;
938 			struct mount *mp = NULL;
939 
940 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
941 				swap_pageouts_ok = 1;
942 			} else {
943 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
944 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
945 				vm_page_count_min());
946 
947 			}
948 
949 			/*
950 			 * We don't bother paging objects that are "dead".
951 			 * Those objects are in a "rundown" state.
952 			 */
953 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
954 				vm_page_unlock(m);
955 				VM_OBJECT_UNLOCK(object);
956 				vm_page_requeue(m);
957 				continue;
958 			}
959 
960 			/*
961 			 * Following operations may unlock
962 			 * vm_page_queue_mtx, invalidating the 'next'
963 			 * pointer.  To prevent an inordinate number
964 			 * of restarts we use our marker to remember
965 			 * our place.
966 			 *
967 			 */
968 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
969 					   m, &marker, pageq);
970 			/*
971 			 * The object is already known NOT to be dead.   It
972 			 * is possible for the vget() to block the whole
973 			 * pageout daemon, but the new low-memory handling
974 			 * code should prevent it.
975 			 *
976 			 * The previous code skipped locked vnodes and, worse,
977 			 * reordered pages in the queue.  This results in
978 			 * completely non-deterministic operation and, on a
979 			 * busy system, can lead to extremely non-optimal
980 			 * pageouts.  For example, it can cause clean pages
981 			 * to be freed and dirty pages to be moved to the end
982 			 * of the queue.  Since dirty pages are also moved to
983 			 * the end of the queue once-cleaned, this gives
984 			 * way too large a weighting to defering the freeing
985 			 * of dirty pages.
986 			 *
987 			 * We can't wait forever for the vnode lock, we might
988 			 * deadlock due to a vn_read() getting stuck in
989 			 * vm_wait while holding this vnode.  We skip the
990 			 * vnode if we can't get it in a reasonable amount
991 			 * of time.
992 			 */
993 			if (object->type == OBJT_VNODE) {
994 				vm_page_unlock_queues();
995 				vm_page_unlock(m);
996 				vp = object->handle;
997 				if (vp->v_type == VREG &&
998 				    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
999 					mp = NULL;
1000 					++pageout_lock_miss;
1001 					if (object->flags & OBJ_MIGHTBEDIRTY)
1002 						vnodes_skipped++;
1003 					vm_page_lock_queues();
1004 					goto unlock_and_continue;
1005 				}
1006 				KASSERT(mp != NULL,
1007 				    ("vp %p with NULL v_mount", vp));
1008 				vm_object_reference_locked(object);
1009 				VM_OBJECT_UNLOCK(object);
1010 				vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1011 				if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
1012 				    curthread)) {
1013 					VM_OBJECT_LOCK(object);
1014 					vm_page_lock_queues();
1015 					++pageout_lock_miss;
1016 					if (object->flags & OBJ_MIGHTBEDIRTY)
1017 						vnodes_skipped++;
1018 					vp = NULL;
1019 					goto unlock_and_continue;
1020 				}
1021 				VM_OBJECT_LOCK(object);
1022 				vm_page_lock(m);
1023 				vm_page_lock_queues();
1024 				/*
1025 				 * The page might have been moved to another
1026 				 * queue during potential blocking in vget()
1027 				 * above.  The page might have been freed and
1028 				 * reused for another vnode.
1029 				 */
1030 				if (m->queue != PQ_INACTIVE ||
1031 				    m->object != object ||
1032 				    TAILQ_NEXT(m, pageq) != &marker) {
1033 					vm_page_unlock(m);
1034 					if (object->flags & OBJ_MIGHTBEDIRTY)
1035 						vnodes_skipped++;
1036 					goto unlock_and_continue;
1037 				}
1038 
1039 				/*
1040 				 * The page may have been busied during the
1041 				 * blocking in vget().  We don't move the
1042 				 * page back onto the end of the queue so that
1043 				 * statistics are more correct if we don't.
1044 				 */
1045 				if (m->busy || (m->oflags & VPO_BUSY)) {
1046 					vm_page_unlock(m);
1047 					goto unlock_and_continue;
1048 				}
1049 
1050 				/*
1051 				 * If the page has become held it might
1052 				 * be undergoing I/O, so skip it
1053 				 */
1054 				if (m->hold_count) {
1055 					vm_page_unlock(m);
1056 					vm_page_requeue(m);
1057 					if (object->flags & OBJ_MIGHTBEDIRTY)
1058 						vnodes_skipped++;
1059 					goto unlock_and_continue;
1060 				}
1061 			}
1062 
1063 			/*
1064 			 * If a page is dirty, then it is either being washed
1065 			 * (but not yet cleaned) or it is still in the
1066 			 * laundry.  If it is still in the laundry, then we
1067 			 * start the cleaning operation.
1068 			 *
1069 			 * decrement page_shortage on success to account for
1070 			 * the (future) cleaned page.  Otherwise we could wind
1071 			 * up laundering or cleaning too many pages.
1072 			 */
1073 			vm_page_unlock_queues();
1074 			if (vm_pageout_clean(m) != 0) {
1075 				--page_shortage;
1076 				--maxlaunder;
1077 			}
1078 			vm_page_lock_queues();
1079 unlock_and_continue:
1080 			vm_page_lock_assert(m, MA_NOTOWNED);
1081 			VM_OBJECT_UNLOCK(object);
1082 			if (mp != NULL) {
1083 				vm_page_unlock_queues();
1084 				if (vp != NULL)
1085 					vput(vp);
1086 				VFS_UNLOCK_GIANT(vfslocked);
1087 				vm_object_deallocate(object);
1088 				vn_finished_write(mp);
1089 				vm_page_lock_queues();
1090 			}
1091 			next = TAILQ_NEXT(&marker, pageq);
1092 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1093 				     &marker, pageq);
1094 			vm_page_lock_assert(m, MA_NOTOWNED);
1095 			continue;
1096 		}
1097 		vm_page_unlock(m);
1098 		VM_OBJECT_UNLOCK(object);
1099 	}
1100 
1101 	/*
1102 	 * Compute the number of pages we want to try to move from the
1103 	 * active queue to the inactive queue.
1104 	 */
1105 	page_shortage = vm_paging_target() +
1106 		cnt.v_inactive_target - cnt.v_inactive_count;
1107 	page_shortage += addl_page_shortage;
1108 
1109 	/*
1110 	 * Scan the active queue for things we can deactivate. We nominally
1111 	 * track the per-page activity counter and use it to locate
1112 	 * deactivation candidates.
1113 	 */
1114 	pcount = cnt.v_active_count;
1115 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1116 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1117 
1118 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1119 
1120 		KASSERT(m->queue == PQ_ACTIVE,
1121 		    ("vm_pageout_scan: page %p isn't active", m));
1122 
1123 		next = TAILQ_NEXT(m, pageq);
1124 		if ((m->flags & PG_MARKER) != 0) {
1125 			m = next;
1126 			continue;
1127 		}
1128 		if (!vm_pageout_page_lock(m, &next)) {
1129 			vm_page_unlock(m);
1130 			m = next;
1131 			continue;
1132 		}
1133 		object = m->object;
1134 		if (!VM_OBJECT_TRYLOCK(object) &&
1135 		    !vm_pageout_fallback_object_lock(m, &next)) {
1136 			VM_OBJECT_UNLOCK(object);
1137 			vm_page_unlock(m);
1138 			m = next;
1139 			continue;
1140 		}
1141 
1142 		/*
1143 		 * Don't deactivate pages that are busy.
1144 		 */
1145 		if ((m->busy != 0) ||
1146 		    (m->oflags & VPO_BUSY) ||
1147 		    (m->hold_count != 0)) {
1148 			vm_page_unlock(m);
1149 			VM_OBJECT_UNLOCK(object);
1150 			vm_page_requeue(m);
1151 			m = next;
1152 			continue;
1153 		}
1154 
1155 		/*
1156 		 * The count for pagedaemon pages is done after checking the
1157 		 * page for eligibility...
1158 		 */
1159 		cnt.v_pdpages++;
1160 
1161 		/*
1162 		 * Check to see "how much" the page has been used.
1163 		 */
1164 		actcount = 0;
1165 		if (object->ref_count != 0) {
1166 			if (m->flags & PG_REFERENCED) {
1167 				actcount += 1;
1168 			}
1169 			actcount += pmap_ts_referenced(m);
1170 			if (actcount) {
1171 				m->act_count += ACT_ADVANCE + actcount;
1172 				if (m->act_count > ACT_MAX)
1173 					m->act_count = ACT_MAX;
1174 			}
1175 		}
1176 
1177 		/*
1178 		 * Since we have "tested" this bit, we need to clear it now.
1179 		 */
1180 		vm_page_flag_clear(m, PG_REFERENCED);
1181 
1182 		/*
1183 		 * Only if an object is currently being used, do we use the
1184 		 * page activation count stats.
1185 		 */
1186 		if (actcount && (object->ref_count != 0)) {
1187 			vm_page_requeue(m);
1188 		} else {
1189 			m->act_count -= min(m->act_count, ACT_DECLINE);
1190 			if (vm_pageout_algorithm ||
1191 			    object->ref_count == 0 ||
1192 			    m->act_count == 0) {
1193 				page_shortage--;
1194 				if (object->ref_count == 0) {
1195 					KASSERT(!pmap_page_is_mapped(m),
1196 				    ("vm_pageout_scan: page %p is mapped", m));
1197 					if (m->dirty == 0)
1198 						vm_page_cache(m);
1199 					else
1200 						vm_page_deactivate(m);
1201 				} else {
1202 					vm_page_deactivate(m);
1203 				}
1204 			} else {
1205 				vm_page_requeue(m);
1206 			}
1207 		}
1208 		vm_page_unlock(m);
1209 		VM_OBJECT_UNLOCK(object);
1210 		m = next;
1211 	}
1212 	vm_page_unlock_queues();
1213 #if !defined(NO_SWAPPING)
1214 	/*
1215 	 * Idle process swapout -- run once per second.
1216 	 */
1217 	if (vm_swap_idle_enabled) {
1218 		static long lsec;
1219 		if (time_second != lsec) {
1220 			vm_req_vmdaemon(VM_SWAP_IDLE);
1221 			lsec = time_second;
1222 		}
1223 	}
1224 #endif
1225 
1226 	/*
1227 	 * If we didn't get enough free pages, and we have skipped a vnode
1228 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1229 	 * if we did not get enough free pages.
1230 	 */
1231 	if (vm_paging_target() > 0) {
1232 		if (vnodes_skipped && vm_page_count_min())
1233 			(void) speedup_syncer();
1234 #if !defined(NO_SWAPPING)
1235 		if (vm_swap_enabled && vm_page_count_target())
1236 			vm_req_vmdaemon(VM_SWAP_NORMAL);
1237 #endif
1238 	}
1239 
1240 	/*
1241 	 * If we are critically low on one of RAM or swap and low on
1242 	 * the other, kill the largest process.  However, we avoid
1243 	 * doing this on the first pass in order to give ourselves a
1244 	 * chance to flush out dirty vnode-backed pages and to allow
1245 	 * active pages to be moved to the inactive queue and reclaimed.
1246 	 */
1247 	if (pass != 0 &&
1248 	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1249 	     (swap_pager_full && vm_paging_target() > 0)))
1250 		vm_pageout_oom(VM_OOM_MEM);
1251 }
1252 
1253 
1254 void
1255 vm_pageout_oom(int shortage)
1256 {
1257 	struct proc *p, *bigproc;
1258 	vm_offset_t size, bigsize;
1259 	struct thread *td;
1260 	struct vmspace *vm;
1261 
1262 	/*
1263 	 * We keep the process bigproc locked once we find it to keep anyone
1264 	 * from messing with it; however, there is a possibility of
1265 	 * deadlock if process B is bigproc and one of it's child processes
1266 	 * attempts to propagate a signal to B while we are waiting for A's
1267 	 * lock while walking this list.  To avoid this, we don't block on
1268 	 * the process lock but just skip a process if it is already locked.
1269 	 */
1270 	bigproc = NULL;
1271 	bigsize = 0;
1272 	sx_slock(&allproc_lock);
1273 	FOREACH_PROC_IN_SYSTEM(p) {
1274 		int breakout;
1275 
1276 		if (PROC_TRYLOCK(p) == 0)
1277 			continue;
1278 		/*
1279 		 * If this is a system, protected or killed process, skip it.
1280 		 */
1281 		if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1282 		    (p->p_pid == 1) || P_KILLED(p) ||
1283 		    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1284 			PROC_UNLOCK(p);
1285 			continue;
1286 		}
1287 		/*
1288 		 * If the process is in a non-running type state,
1289 		 * don't touch it.  Check all the threads individually.
1290 		 */
1291 		breakout = 0;
1292 		FOREACH_THREAD_IN_PROC(p, td) {
1293 			thread_lock(td);
1294 			if (!TD_ON_RUNQ(td) &&
1295 			    !TD_IS_RUNNING(td) &&
1296 			    !TD_IS_SLEEPING(td)) {
1297 				thread_unlock(td);
1298 				breakout = 1;
1299 				break;
1300 			}
1301 			thread_unlock(td);
1302 		}
1303 		if (breakout) {
1304 			PROC_UNLOCK(p);
1305 			continue;
1306 		}
1307 		/*
1308 		 * get the process size
1309 		 */
1310 		vm = vmspace_acquire_ref(p);
1311 		if (vm == NULL) {
1312 			PROC_UNLOCK(p);
1313 			continue;
1314 		}
1315 		if (!vm_map_trylock_read(&vm->vm_map)) {
1316 			vmspace_free(vm);
1317 			PROC_UNLOCK(p);
1318 			continue;
1319 		}
1320 		size = vmspace_swap_count(vm);
1321 		vm_map_unlock_read(&vm->vm_map);
1322 		if (shortage == VM_OOM_MEM)
1323 			size += vmspace_resident_count(vm);
1324 		vmspace_free(vm);
1325 		/*
1326 		 * if the this process is bigger than the biggest one
1327 		 * remember it.
1328 		 */
1329 		if (size > bigsize) {
1330 			if (bigproc != NULL)
1331 				PROC_UNLOCK(bigproc);
1332 			bigproc = p;
1333 			bigsize = size;
1334 		} else
1335 			PROC_UNLOCK(p);
1336 	}
1337 	sx_sunlock(&allproc_lock);
1338 	if (bigproc != NULL) {
1339 		killproc(bigproc, "out of swap space");
1340 		sched_nice(bigproc, PRIO_MIN);
1341 		PROC_UNLOCK(bigproc);
1342 		wakeup(&cnt.v_free_count);
1343 	}
1344 }
1345 
1346 /*
1347  * This routine tries to maintain the pseudo LRU active queue,
1348  * so that during long periods of time where there is no paging,
1349  * that some statistic accumulation still occurs.  This code
1350  * helps the situation where paging just starts to occur.
1351  */
1352 static void
1353 vm_pageout_page_stats()
1354 {
1355 	vm_object_t object;
1356 	vm_page_t m,next;
1357 	int pcount,tpcount;		/* Number of pages to check */
1358 	static int fullintervalcount = 0;
1359 	int page_shortage;
1360 
1361 	page_shortage =
1362 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1363 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1364 
1365 	if (page_shortage <= 0)
1366 		return;
1367 
1368 	vm_page_lock_queues();
1369 	pcount = cnt.v_active_count;
1370 	fullintervalcount += vm_pageout_stats_interval;
1371 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1372 		tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1373 		    cnt.v_page_count;
1374 		if (pcount > tpcount)
1375 			pcount = tpcount;
1376 	} else {
1377 		fullintervalcount = 0;
1378 	}
1379 
1380 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1381 	while ((m != NULL) && (pcount-- > 0)) {
1382 		int actcount;
1383 
1384 		KASSERT(m->queue == PQ_ACTIVE,
1385 		    ("vm_pageout_page_stats: page %p isn't active", m));
1386 
1387 		next = TAILQ_NEXT(m, pageq);
1388 		if ((m->flags & PG_MARKER) != 0) {
1389 			m = next;
1390 			continue;
1391 		}
1392 		vm_page_lock_assert(m, MA_NOTOWNED);
1393 		if (!vm_pageout_page_lock(m, &next)) {
1394 			vm_page_unlock(m);
1395 			m = next;
1396 			continue;
1397 		}
1398 		object = m->object;
1399 		if (!VM_OBJECT_TRYLOCK(object) &&
1400 		    !vm_pageout_fallback_object_lock(m, &next)) {
1401 			VM_OBJECT_UNLOCK(object);
1402 			vm_page_unlock(m);
1403 			m = next;
1404 			continue;
1405 		}
1406 
1407 		/*
1408 		 * Don't deactivate pages that are busy.
1409 		 */
1410 		if ((m->busy != 0) ||
1411 		    (m->oflags & VPO_BUSY) ||
1412 		    (m->hold_count != 0)) {
1413 			vm_page_unlock(m);
1414 			VM_OBJECT_UNLOCK(object);
1415 			vm_page_requeue(m);
1416 			m = next;
1417 			continue;
1418 		}
1419 
1420 		actcount = 0;
1421 		if (m->flags & PG_REFERENCED) {
1422 			vm_page_flag_clear(m, PG_REFERENCED);
1423 			actcount += 1;
1424 		}
1425 
1426 		actcount += pmap_ts_referenced(m);
1427 		if (actcount) {
1428 			m->act_count += ACT_ADVANCE + actcount;
1429 			if (m->act_count > ACT_MAX)
1430 				m->act_count = ACT_MAX;
1431 			vm_page_requeue(m);
1432 		} else {
1433 			if (m->act_count == 0) {
1434 				/*
1435 				 * We turn off page access, so that we have
1436 				 * more accurate RSS stats.  We don't do this
1437 				 * in the normal page deactivation when the
1438 				 * system is loaded VM wise, because the
1439 				 * cost of the large number of page protect
1440 				 * operations would be higher than the value
1441 				 * of doing the operation.
1442 				 */
1443 				pmap_remove_all(m);
1444 				vm_page_deactivate(m);
1445 			} else {
1446 				m->act_count -= min(m->act_count, ACT_DECLINE);
1447 				vm_page_requeue(m);
1448 			}
1449 		}
1450 		vm_page_unlock(m);
1451 		VM_OBJECT_UNLOCK(object);
1452 		m = next;
1453 	}
1454 	vm_page_unlock_queues();
1455 }
1456 
1457 /*
1458  *	vm_pageout is the high level pageout daemon.
1459  */
1460 static void
1461 vm_pageout()
1462 {
1463 	int error, pass;
1464 
1465 	/*
1466 	 * Initialize some paging parameters.
1467 	 */
1468 	cnt.v_interrupt_free_min = 2;
1469 	if (cnt.v_page_count < 2000)
1470 		vm_pageout_page_count = 8;
1471 
1472 	/*
1473 	 * v_free_reserved needs to include enough for the largest
1474 	 * swap pager structures plus enough for any pv_entry structs
1475 	 * when paging.
1476 	 */
1477 	if (cnt.v_page_count > 1024)
1478 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1479 	else
1480 		cnt.v_free_min = 4;
1481 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1482 	    cnt.v_interrupt_free_min;
1483 	cnt.v_free_reserved = vm_pageout_page_count +
1484 	    cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1485 	cnt.v_free_severe = cnt.v_free_min / 2;
1486 	cnt.v_free_min += cnt.v_free_reserved;
1487 	cnt.v_free_severe += cnt.v_free_reserved;
1488 
1489 	/*
1490 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1491 	 * that these are more a measure of the VM cache queue hysteresis
1492 	 * then the VM free queue.  Specifically, v_free_target is the
1493 	 * high water mark (free+cache pages).
1494 	 *
1495 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1496 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1497 	 * be big enough to handle memory needs while the pageout daemon
1498 	 * is signalled and run to free more pages.
1499 	 */
1500 	if (cnt.v_free_count > 6144)
1501 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1502 	else
1503 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1504 
1505 	if (cnt.v_free_count > 2048) {
1506 		cnt.v_cache_min = cnt.v_free_target;
1507 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1508 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1509 	} else {
1510 		cnt.v_cache_min = 0;
1511 		cnt.v_cache_max = 0;
1512 		cnt.v_inactive_target = cnt.v_free_count / 4;
1513 	}
1514 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1515 		cnt.v_inactive_target = cnt.v_free_count / 3;
1516 
1517 	/* XXX does not really belong here */
1518 	if (vm_page_max_wired == 0)
1519 		vm_page_max_wired = cnt.v_free_count / 3;
1520 
1521 	if (vm_pageout_stats_max == 0)
1522 		vm_pageout_stats_max = cnt.v_free_target;
1523 
1524 	/*
1525 	 * Set interval in seconds for stats scan.
1526 	 */
1527 	if (vm_pageout_stats_interval == 0)
1528 		vm_pageout_stats_interval = 5;
1529 	if (vm_pageout_full_stats_interval == 0)
1530 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1531 
1532 	swap_pager_swap_init();
1533 	pass = 0;
1534 	/*
1535 	 * The pageout daemon is never done, so loop forever.
1536 	 */
1537 	while (TRUE) {
1538 		/*
1539 		 * If we have enough free memory, wakeup waiters.  Do
1540 		 * not clear vm_pages_needed until we reach our target,
1541 		 * otherwise we may be woken up over and over again and
1542 		 * waste a lot of cpu.
1543 		 */
1544 		mtx_lock(&vm_page_queue_free_mtx);
1545 		if (vm_pages_needed && !vm_page_count_min()) {
1546 			if (!vm_paging_needed())
1547 				vm_pages_needed = 0;
1548 			wakeup(&cnt.v_free_count);
1549 		}
1550 		if (vm_pages_needed) {
1551 			/*
1552 			 * Still not done, take a second pass without waiting
1553 			 * (unlimited dirty cleaning), otherwise sleep a bit
1554 			 * and try again.
1555 			 */
1556 			++pass;
1557 			if (pass > 1)
1558 				msleep(&vm_pages_needed,
1559 				    &vm_page_queue_free_mtx, PVM, "psleep",
1560 				    hz / 2);
1561 		} else {
1562 			/*
1563 			 * Good enough, sleep & handle stats.  Prime the pass
1564 			 * for the next run.
1565 			 */
1566 			if (pass > 1)
1567 				pass = 1;
1568 			else
1569 				pass = 0;
1570 			error = msleep(&vm_pages_needed,
1571 			    &vm_page_queue_free_mtx, PVM, "psleep",
1572 			    vm_pageout_stats_interval * hz);
1573 			if (error && !vm_pages_needed) {
1574 				mtx_unlock(&vm_page_queue_free_mtx);
1575 				pass = 0;
1576 				vm_pageout_page_stats();
1577 				continue;
1578 			}
1579 		}
1580 		if (vm_pages_needed)
1581 			cnt.v_pdwakeups++;
1582 		mtx_unlock(&vm_page_queue_free_mtx);
1583 		vm_pageout_scan(pass);
1584 	}
1585 }
1586 
1587 /*
1588  * Unless the free page queue lock is held by the caller, this function
1589  * should be regarded as advisory.  Specifically, the caller should
1590  * not msleep() on &cnt.v_free_count following this function unless
1591  * the free page queue lock is held until the msleep() is performed.
1592  */
1593 void
1594 pagedaemon_wakeup()
1595 {
1596 
1597 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1598 		vm_pages_needed = 1;
1599 		wakeup(&vm_pages_needed);
1600 	}
1601 }
1602 
1603 #if !defined(NO_SWAPPING)
1604 static void
1605 vm_req_vmdaemon(int req)
1606 {
1607 	static int lastrun = 0;
1608 
1609 	mtx_lock(&vm_daemon_mtx);
1610 	vm_pageout_req_swapout |= req;
1611 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1612 		wakeup(&vm_daemon_needed);
1613 		lastrun = ticks;
1614 	}
1615 	mtx_unlock(&vm_daemon_mtx);
1616 }
1617 
1618 static void
1619 vm_daemon()
1620 {
1621 	struct rlimit rsslim;
1622 	struct proc *p;
1623 	struct thread *td;
1624 	struct vmspace *vm;
1625 	int breakout, swapout_flags;
1626 
1627 	while (TRUE) {
1628 		mtx_lock(&vm_daemon_mtx);
1629 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1630 		swapout_flags = vm_pageout_req_swapout;
1631 		vm_pageout_req_swapout = 0;
1632 		mtx_unlock(&vm_daemon_mtx);
1633 		if (swapout_flags)
1634 			swapout_procs(swapout_flags);
1635 
1636 		/*
1637 		 * scan the processes for exceeding their rlimits or if
1638 		 * process is swapped out -- deactivate pages
1639 		 */
1640 		sx_slock(&allproc_lock);
1641 		FOREACH_PROC_IN_SYSTEM(p) {
1642 			vm_pindex_t limit, size;
1643 
1644 			/*
1645 			 * if this is a system process or if we have already
1646 			 * looked at this process, skip it.
1647 			 */
1648 			PROC_LOCK(p);
1649 			if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1650 				PROC_UNLOCK(p);
1651 				continue;
1652 			}
1653 			/*
1654 			 * if the process is in a non-running type state,
1655 			 * don't touch it.
1656 			 */
1657 			breakout = 0;
1658 			FOREACH_THREAD_IN_PROC(p, td) {
1659 				thread_lock(td);
1660 				if (!TD_ON_RUNQ(td) &&
1661 				    !TD_IS_RUNNING(td) &&
1662 				    !TD_IS_SLEEPING(td)) {
1663 					thread_unlock(td);
1664 					breakout = 1;
1665 					break;
1666 				}
1667 				thread_unlock(td);
1668 			}
1669 			if (breakout) {
1670 				PROC_UNLOCK(p);
1671 				continue;
1672 			}
1673 			/*
1674 			 * get a limit
1675 			 */
1676 			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1677 			limit = OFF_TO_IDX(
1678 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1679 
1680 			/*
1681 			 * let processes that are swapped out really be
1682 			 * swapped out set the limit to nothing (will force a
1683 			 * swap-out.)
1684 			 */
1685 			if ((p->p_flag & P_INMEM) == 0)
1686 				limit = 0;	/* XXX */
1687 			vm = vmspace_acquire_ref(p);
1688 			PROC_UNLOCK(p);
1689 			if (vm == NULL)
1690 				continue;
1691 
1692 			size = vmspace_resident_count(vm);
1693 			if (limit >= 0 && size >= limit) {
1694 				vm_pageout_map_deactivate_pages(
1695 				    &vm->vm_map, limit);
1696 			}
1697 			vmspace_free(vm);
1698 		}
1699 		sx_sunlock(&allproc_lock);
1700 	}
1701 }
1702 #endif			/* !defined(NO_SWAPPING) */
1703